SD3.5M-SOLACE-on-FlowGRPO-PickScore

LoRA adapter from SOLACE (Self-cOnfidence reward for aLigning text-to-imAge models via ConfidencE optimization), CVPR 2026.

SOLACE self-confidence post-training applied on top of a Flow-GRPO model that was post-trained on the PickScore human-preference reward.

• Base model: stabilityai/stable-diffusion-3.5-medium
• Training pipeline: SD3.5-Medium → Flow-GRPO post-training → SOLACE self-confidence post-training
• Code: https://github.com/wookiekim/SOLACE
• Adapter type: PEFT LoRA (rank 32, applied to the SD3 transformer attention projections)

Usage

import torch
from diffusers import StableDiffusion3Pipeline
from peft import PeftModel

model_id = "stabilityai/stable-diffusion-3.5-medium"
lora_ckpt_path = "wookiekim/SD3.5M-SOLACE-on-FlowGRPO-PickScore"
device = "cuda"

# Load base model and apply the SOLACE LoRA adapter
pipe = StableDiffusion3Pipeline.from_pretrained(model_id, torch_dtype=torch.float16)
pipe.transformer = PeftModel.from_pretrained(pipe.transformer, lora_ckpt_path)
pipe.transformer = pipe.transformer.merge_and_unload()
pipe = pipe.to(device)

prompt = "a photo of a cat wearing a small red hat"
image = pipe(
    prompt,
    height=512,
    width=512,
    num_inference_steps=40,
    guidance_scale=4.5,
    negative_prompt="",
).images[0]
image.save("solace.png")

Note: This adapter already contains the combined Flow-GRPO + SOLACE update as a single LoRA — load it directly on the base SD3.5-Medium model; no separate Flow-GRPO adapter is required.

Citation

@inproceedings{kim2026solace,
    title={Improving Text-to-Image Generation with Intrinsic Self-Confidence Rewards},
    author={Kim, Wookyoung and others},
    booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    year={2026}
}

Acknowledgments

This work builds upon Flow-GRPO by Jie Liu et al.